8
$\begingroup$

I’m starting to see more and more answers from brand-new users that appear to be ChatGPT-driven copies. They have a veneer of truth but otherwise often miss the mark. Or, if they do fully address the question in the title, I’m bothered by the approach - mostly just the absence of any disclosure or disclaimer that the answer was provided by ChatGPT.

I’m flagging them and downvoting when I see them. I’ve provided some (mostly moderate and hopefully polite) comments to state that this not acceptable practice and should be retrained. But I’m losing patience.

I might be willing to take a softer line if there were any disclaimer from the OP that the answer was guided by ChatGPT. Or, leaving the answer as-is but marking it as CW.

Also my Turing test might flag some false-negatives - but we have some electronic tools for that.

It seems like the red-flags are:

  • New users
  • Hot questions
  • Simple questions without a lot of background that can be put into ChatGPT prompt
  • The answer is way longer than the question and lacks any pointers to other info or papers, etc.
  • No grammatical error or typos

I’m looking for a consensus. Should we remove these at will, or flag, or take some other course of action?

$\endgroup$
1
  • 2
    $\begingroup$ Interesting, I didn't notice this (although I don't check a lot of questions). Do you have an example? $\endgroup$
    – Rammus
    Apr 19 at 15:20

1 Answer 1

Reset to default
5
$\begingroup$

To be honest, I personally thought it was a given that we should not accept chatGPT-generated posts. But it's not a network-wide policy so it's good to have this discussion.

In fact, we already had a number of such posts, which were promptly removed (and in some cases where the user was posting a large number of these answers network-wide, they were banned). The most recent example, and the one that presumably spurred this question, is https://quantumcomputing.stackexchange.com/a/32221/55 (can be seen only if you have high enough rep). I quickly went through the last 300 deleted answers, dating from 31/08/2021 to now, so given the release date of chatGPT these should be all. I only looked at the first or first sentences in the post to figure out whether the post was likely to be chatGPT generated or not. Amusingly, this worked 100% of the time (or at least, all chatGPT-looking first sentences corresponded to posts deleted as chatGPT generated). Here's what I found:

Most of these answers were deleted in response to a user flag, to which we agreed.

I think we should hard delete all such questions or answers. Regarding the issue of figuring out whether a post is actually chatGPT-generated, that's actually not particularly hard. You can usually tell by looking (if you've played with chatGPT enough I guess), or you can use online tools such as https://detectgpt.com/ to get further confirmation. They might not be 100% accurate, but I'd say common sense plus online tools give us a high enough confidence to act on it. I can very well imagine a future where future iterations of these tools won't be so easy to detect, but that's going to be a discussion for another day.

As for the reasons why I don't think we should allow these posts, I feel like this was discussed in-length network-wide, so there's no need to go through it again here. I already linked above the mother meta discussion. You might also have a look at the discussion on stackoverflow. Another one perhaps closer to our case is the one on physics.SE. As discussed in length there, chatGPT-generated posts are at best useless, and usually harmful because they look fine, but might hide several imprecisions in a nice word cloud.

$\endgroup$
8
  • $\begingroup$ "it doesn't seem like we can see a list of deleted posts" For users that have enough rep, you can see the deleted posts over the last 30 days here: quantumcomputing.stackexchange.com/… and for diamond moderators there's other tools available. When I execute this search: quantumcomputing.stackexchange.com/search?q=deleted%3A1 I only get my deleted posts, but maybe when you search it you will find all deleted posts. You can check the teams page for diamond moderators, if you want to find tools for finding deleted posts. $\endgroup$
    – user1271772
    Apr 19 at 16:41
  • $\begingroup$ About the recently deleted answer to which you provided a link here, I was one of the people that flagged it, but I wonder if in a comment on the deleted answer, you could provide to us a link to what led you to conclude that it's a ChatGPT answer? For example did you use zerogpt? In my flag I said that it "looks" like it's from ChatGPT, but I hadn't found "proof" yet. $\endgroup$
    – user1271772
    Apr 19 at 16:45
  • $\begingroup$ I (MarkS) didn’t run the recent answer through those tools mentioned by gIS for this case; I relied on my own (potentially fallible) Turing test this time. But I have run those tools otherwise, and own accuracy isn’t bad. $\endgroup$
    – Mark Spinelli
    Apr 19 at 17:34
  • 1
    $\begingroup$ I’m also interested when ChatGP’s answer is not bad, or at least somewhat informative. Perhaps don’t flag if the OP provides a disclaimer saying it was GPT inspired? And/or mark as CW so there’s no reputation farming? We all know that GPT scored a low B on Aaronson’s final. $\endgroup$
    – Mark Spinelli
    Apr 19 at 17:50
  • 2
    $\begingroup$ @user1271772 ah, good thinking. I used the link to quickly go through the last deleted answers and retrieve the ones deleted for being chatGPT (well, at least the ones that seemed like they were from the first few sentences you get from the search page). About the last answer at hand, the zerogpt tool linked above gives it as chatGPT-generated with "100% confidence". $\endgroup$
    – glS Mod
    Apr 19 at 18:30
  • $\begingroup$ @MarkS I mean, if it was "chatGPT-inspired" that means it was written by the user anyway, so it probably won't raise flags when looked at. Also, if an answer where chatGPT was involved to whatever degree is genuinely well-written, meaningful, and to the point, I feel like we wouldn't even notice such involvement, so again it won't raise any red flags. I myself sometimes use it to improve writing style and things like that. It's a great tool, if you are able to assess its output's correctness $\endgroup$
    – glS Mod
    Apr 19 at 18:34
  • $\begingroup$ all the examples of posts deleted for being chatGPT generated are effectively bad answers formatted in a strangely good way, but often hide nonsense, or are bad answers where the user is not able to give feedback about questions asked in the comments $\endgroup$
    – glS Mod
    Apr 19 at 18:36
  • $\begingroup$ Right I don’t think we’ve done anything wrong yet. A flag by a user with xp to flag meets the prima facie expectation of being GPT generated and the burden shifts to the poster to provide evidence otherwise. Perhaps this is something reviewers can be aware of for reviewing late answers or first answers etc. For hot questions I also try to protect after any suspect answer. $\endgroup$
    – Mark Spinelli
    Apr 19 at 19:58

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .